Data-Driven Inertial Navigation assisted by 5G UL-TDoA Positioning

The increasing availability of 5G-enabled smartphones will enable new possibilities in the context of indoor positioning. 5G accurate positioning could be considered the next big development for GNSS-challenged navigation. Experimental projects for 5G based positioning already have started, but the research is yet missing promising results in real life scenarios with real 5G-network positioning. Recently, the interest in training deep neural networks (DNNs) using modern smartphone Inertial Measurement Units (IMUs) for location estimation has also been growing. However, the lack of diverse quality labelled data for training and evaluating has limited the adoption of DNNs in sensor-based positioning tasks, especially for real life and long trajectory estimation. In this paper, we show how real Uplink Time Difference of Arrival (UL-TDoA) 5G positioning data can be used in conjunction with real-time inertial navigation systems in commercial smartphones for indoor scenarios. Moreover, a data driven inertial navigation approach, utilizing DNNs based on different qualities of label data including 5G outputs has been developed. Qualitative and quantitative evaluations of the competing methods over three inertial navigation benchmarks in comparison with the merged dataset including our 5G experiments has been done. As the result of this work, we could show that long-time and real-life indoor positioning on a commercial smartphone is feasible even in heterogenous 5G-positioning service areas. This has been realized by: 1) The realization of a 5G UL-TDoA positioning followed by an extended experiment in a largescale environment. 2) The development of a streamlined state estimation approach which builds upon DNN based odometry outputs and is optimized by the reference 5G absolute positions. 3) The development of a web and a smartphone application used for real-life navigation and research data collection up to the result analysis. Based on our accuracy analysis using Mean Absolut Error (MAE) and Circular Error Probability (CEP) metric values, it can be stated that a sub meter accuracy in combination of 5G positioning DNNs can be achieved for real-life indoor positioning scenarios.